Recommendation algorithm based on user score probability and project type

The core idea of the recommendation algorithm is to obtain the information implied in the user’s behavior, identify the user’s behavior, use the collective wisdom [4] to match the user, and recommend the product to the user.

The traditional recommendation algorithm only pays attention to the value of the product scored by the user [5, 6], ignoring the user’s implicit information in the behavior of scoring the commodities. The traditional algorithms indicate that the user randomly selects some of the commodities and scores them according to the degree of preference for the commodities. A high score shows that the user likes this commodity, and the low score shows that the user does not like the commodities. In the era of online shopping information explosion, a user’s shopping behavior is based on their own needs and preferences [7‐10]. The user will score the goods according to the quality of commodities, customer service attitude, logistics speed, and other factors. If the user is not interested in a commodity, he will not buy it. That is to say, giving a commodity a low score can only indicate that the user is not satisfied with this product. In the traditional algorithm, this dissatisfaction is spread to the same type of commodity, making the system think that the user’s preference for similar products is reduced and affecting the system’s recommendation for similar products. Therefore, it is a subjective behavior of the user to select a product and score it and this behavior is an invisible embodiment of the user’s interest preference [11‐13].

About users’ behavior, there are some different views between this paper and the traditional algorithms:

For example, as shown in Fig. 1, user A loves to watch action movies, but does not love science fiction and comedy. As the number of movies on the video site is very large, so user A will choose his favorite movie to watch. First of all, user A will choose the action movie on the video site, then scores it after watching the movie. Because they do not like to see the comedy and science fiction film, user A cannot evaluate such films. User A finds a movie called “IP MAN” [18, 19] in the action movie category. After watching it, user A feels that the clarity of the picture is not good, so he gives the film a low score. If user A does not like the action movie, he will not see this category of movies. The traditional collaborative filter [20, 21] considers that as long as the user gives a low score that users do not like this type of film, but the fact is that if the user does not like a film, users will not pay attention to this type, let alone watch and to score.

This paper believes that the score value cannot indicate whether the user likes this kind of commodity or not, but only indicate that the user is not satisfied with the current product. Users pay attention to and want to consume the products that they are interested in. Users will score a high score for products that they are interested in and satisfied with, and low scores for products that they are interested in but not satisfied with. Therefore, regardless of whether the user gives a product a high or low score, the behavior of the user to score the product fully indicates that the user is interested in this type of product.

Based on the above viewpoints and the implicit expression of the user score behavior [22, 23], this paper designs a recommendation algorithm model based on user score preference and project type. By analyzing the user score behavior, our algorithm obtains the user’s preference for the commodity. According to the preference of the user, we can predict the probability of the user to score the target commodities. The recommendation system combines the similarity calculated from the score value with the similarity calculated from the user score probability and the project type to make the similarity between the users more accurate. The similarity calculation of recommendation algorithm framework based on score preference and the project type is shown in Fig. 2.

As shown in Fig. 2, the behavior that a user gives a score to a commodity is called the user rating behavior. The score value that the user gives to the commodity is called the user rating value. The frame calculates the score probability P based on the user rating behavior and calculates the similarity S2 according to the score probability P and the project type. Then, we calculate the similarity S1 according to the user rating value. Finally, we combine the two similarity values S1 and S2 to obtain the ultimate similarity S [24, 25].

On the basis of the conventional algorithm based on neighborhood, the ideas of the user behavior in Fig. 1 and the improved similarity of Fig. 2 are integrated. The improved algorithm framework of this chapter is obtained, which is called the recommendation algorithm framework of scoring preference and project type. This framework makes full use of the user’s preference information and calculates the user interest in a kind of commodity, that is, the probability of scoring. We will get the scoring probability and improved similarity by calculating [26]. The process of the model of the score preference and the project type is shown in Fig. 3.

The recommendation algorithm based on the user score preference and project type combines the probability of the user to score the commodity with the user’s prediction value to the product. Some studies have shown that making full use of the user’s behavior of scoring the product can effectively improve the recommendation accuracy of recommendation algorithm. The model takes full advantage of the user’s scoring behavior, excavates the user’s implied hobbies, and predicts the product type that the user may be interested in.

The core idea of the model based on the user score preference and project types is shown in Fig. 3: Firstly, we calculate the similarity S1 based on the matrix. Then, we calculate the possibility Pro of the user to score the product according to the preference information implied in the user rating behavior. The second similarity S2 is calculated according to Pro and the type of the product. Due to the different weights of the two similarities, the final similarity of the user S is obtained by combining the two similarities S1 and S2 [27, 28]. For all users, the similarity between any two forms an M×M similarity matrix (M is the number of users).

For example, in order to calculate the similarity of user A and user B, the algorithm reads the scoring information from the data set, gets the score matrix of 943×1682 (943 represents 943 users, and 1682 represents 1682 commodities), and calculates the Sim1(A,B) by score matrix and Pearson’s formula. Then, a 943×18 score count matrix and a 943×18 score probability matrix are created (943 represents 943 users, and 18 represents of 18 types). Next, the algorithm traverses the score matrix, records the number of user score for each commodity into the 943×18 score matrix. The algorithm also traverses the score count matrix, calculates the probability of user score for each commodity, and records it into the scoring probability matrix. We obtain the second similarity S2(A,B) through probability matrix and Pearson’s formula and get the final S(A,B) by combining S1(A, B) and S2(A, B). By calculating the similarity by the above way, we can obtain a 943×943 similarity matrix.

The UPCF algorithm takes full advantage of the user’s behavior and the type of information of the product to score the product, which is the main difference between the UPCF algorithm and the traditional recommendation algorithm. The two-step prediction recommendation algorithm proposed in [3] and the probabilistic latent semantic recommendation algorithm based on autonomous prediction proposed in [13, 29] make full use of the user’s behavior information to score the product. The difference between them is that the UPCF algorithm uses a different approach when calculating the score probability and considers the type of information of the project when calculating the similarity.

In order to verify the effectiveness of the framework, this paper combines IBCF with the framework in Fig. 2 to propose an algorithm of fusing score preferences and project types. The next section will detail the UPCF algorithm.

If the value is not 0, the user has scored the commodity; otherwise, there is no score on the commodities. Traverse the target user’s n-dimensional score vector, count the type of commodities and the number of times each commodity is scored, and put the statistic results into the list. Each item in the list is an <i, n> binary relationship group, where i is the commodity type, and n is the number of times the commodities have been scored. We predict the users’ interest in this type of product according to the number of users scoring a certain type in the list, that is, predicting the probability of users to score the commodity. If the target commodity type is j, N(j) represents the scoring number of u on the j type and M is the total number of the user to score commodities. The user score probability is calculated as follows:

The specific implementation of the score probability prediction is shown in the following pseudo code:

The cosine similar and Pearson et al. [30‐32] are the most common way to calculate similarity in the conventional collaborative filtering algorithms. But regardless of the kind of calculation, the database of the calculation is the commodities’ score commonly used by the users. This calculation ignores the commodity type [33]. This section incorporates the commodity type and the scoring probability of the user on the basis of the traditional similarity calculation method and adjusts the ratio of the two similarities. The improved similarity formula is as follows:

where Sim(U_i, U_j) is the final similarity, S(U_i, U_j) is the similarity calculated by using the user’s score value, S_sort(U_i, U_j) is the similarity calculated by the commodity type and the scoring probability of product. The formula to calculate S_sort(U_i, U_j) is as follows:

where L(U_i) is a type collection of commodities that are scored by U_i. L(U_j) is a type collection of commodities that are scored by U_j. \( {P}_{U_ik} \) is the scoring probability of U_i for the k type, and \( {\overline{P}}_{U_i} \) is the average of the scoring probability of U_i for all the types. The k type is one of the intersection types scored by U_i and U_j.

The selection of the nearest neighbor and the scoring prediction and scoring criteria have been described in detail in the previous section, and it is no longer described here.

There are two conditions for the target user to choose the nearest neighbor [34‐37]. Firstly, the selected neighbor is highly similar to the target user. Secondly, the selected neighbor has been scored on the target commodity.

When selecting the best neighbor, there is a threshold needed to set in order to prevent the existence of less similar individuals that affects the final results in collaborative filtering. Only the neighbor who has given the target product score and the similarity is greater than the threshold value that can become the target neighbor. The selection of neighbors is as follows:

where KN(U_m) is the neighbor list of the user U_m and β is the threshold, which can be set to the average value of the similarity of all the users who are similar to the user U_m. The specific implementation of selecting neighbor is as follows:

In the calculation of the predicted score, the traditional collaborative filtering algorithm only focuses on the similarity [38, 39] between the neighbor and the target user and the neighbor’s score for the prediction item. Each user has different scoring criteria. For example, some users give three points to show that they like that product, and some users need to give five points to express the same meaning. In order to solve this problem, this algorithm takes the average value of the user’s score into account to resolve the difference between users. The user’s rating for item v is

where \( f=\exp \left\{-1+\alpha \left({\overline{r}}_{ui}-{\overline{r}}_{uj}\right)\right\} \), exp represents the exponential function based on e, \( {\overline{r}}_{ui} \) is the average of U_i, \( {r}_{u_jv} \) is score of U_j to v, α is the attenuation factor, and Score(U_i, v) is the prediction score of U_i to v.

The evaluation indicators of the recommendation system can be summarized as accuracy and the other indicators out of accuracy [40, 41]. The accuracy of this paper is mainly referring to the accuracy index of the prediction score. This kind of indicator is to judge the accuracy by comparing the difference between the prediction score and the real score. The most commonly used is the MAE (mean absolute error), |test| is the test set, r_uv is a prediction of U to V, and \( {r}_{uv}^{\mathrm{test}} \) is the real score of U to V in the test set.

MAE is calculated as follows:

MAE is easy to understand and to calculate, but it also has some shortcomings that the MAE makes a contribution to the inaccurate prediction of low-score products. The RMSE (root mean square error) is also an evaluation indicator related to MAE; RMSE is calculated as follows:

In RMSE, each absolute error is squared, which makes the larger absolute error becomes larger.

In order to verify the validity of the collaborative filtering recommendation algorithm of score preference and project type, the experiment was validated on the MovieLens set provided by GroupLens. The MovieLens data set is collected by the GroupLens Study Group of the University of Minnesota [42‐44], which contains three different versions. This chapter selects the ml-100K data set for experimentation. The data set has 943 users and 1682 movies and 943×1682 score records. The score is 0 or a positive integer between 1 and 5, the score is 0 that the user did not score the product, the higher the score indicates that the higher the degree of the user’s preferences for commodities. Ninety percent of the data set was randomly selected as the training set, and the rest was used for the experiment.

The acquisition of the experimental data samples is from the MovieLens data set [45‐47] provided by GroupLens. We select the ml-100K version of MovieLens. The traditional collaborative filtering algorithm based on the nearest neighbor, GSCF algorithm, and UPCF algorithm is run on the data sets train1 and test1 and data sets train2 and test2, and then compare and analyze the difference between the MAE value and RMSE value of the three algorithms [48‐50]. According to the analysis of the three algorithms recommendation results, the specific steps of the experimental design are as follows:

The first step, the division of the data set: the data set will be divided into two parts according to the proportion of 9:1 in accordance with the principle of completely random. One class is called the training set train, and the less is called the test set test. The ml-100K data set is divided into several times, and we obtain the training sets train1, train2, train3 and so on, as well as the test set corresponding to the training sets test1, test2, test3 and so on.

The second step, the prediction of scoring probability: the probability of the user’s score for each type is calculated according to Section 3.1, and a score probability matrix of 943×18 is obtained. Nine hundred forty-three lines represent 943 users, and 18 columns represent 18 types of movies.

The third step, the similarity calculation: the similarity S1 is calculated by the score matrix, and the similarity S2 is calculated by the score probability. We will get a similarity matrix by combining the two similarities; the similarity matrix S is as follows:

\( S=\left[\begin{array}{cccc}{S}_{11}& {S}_{12}& \dots & {S}_{1m}\\ {}{S}_{21}& \dots & \dots & \dots \\ {}\dots & \dots & \dots & \dots \\ {}{S}_{m1}& \dots & \dots & {S}_{mm}\end{array}\right] \)

S_m1 is the similarity between user m and user 1 in the matrix.

The fourth step is to choose the nearest neighbor according to the similarity.

The fifth step is to calculate the scoring error MAE, RMSE.

This section is compared with the traditional user-based propulsion algorithm UBCF, traditional item-based algorithm IBCF, a recommendation algorithm GSCF based on graph structure and project type. For the UPCF algorithm, we carry out experimental analysis. By comparing the influence of different neighbors on the MAE and RMSE of these four recommendation algorithms, the number of neighbors is the same, the difference between the MAE value and the RMSE value of the four recommendation algorithms is obtained. The number of neighbors which selected 10 to 80 variables was shown in the data sets train1 and test1 and data sets train2 and test2 of these four algorithms (UBCF, IBCF, GSCF, UPCF) in the MAE and RMSE performance. In the data sets train1 and test1, the three recommendation algorithm’s MAE value of the comparison is shown in Table 1.

In Table 1, UBCF is based on the user’s algorithm, IBCF is a conventional item-based algorithm, GSCF is an algorithm based on graph structure and project type, and UPCF is an algorithm based on user score preference and project type. When the number of neighbors is the same, the MAE value of the improved algorithm UPCF is the smallest, that is, the prediction error is the smallest and the recommendation performance is the best. When the number of neighbors is different, the MAE value of the four algorithms decreases first and then increases with the growth of the neighbors. It shows that the MAE is affected by the neighbors, in other words, the performance is affected by the neighbors. In order to make the comparison of the four algorithms more obvious, this chapter draws the MAE values into line graphs, as shown in Fig. 4.

We take the number of neighbors as variables; with the increase of the number of neighbors, the MAE values of the four algorithms are reduced first and then flattened. This shows that the number of neighbors has a certain impact on the scoring error, when the number of neighbors is enough, this effect gradually weakened. In the case of the same number of neighbors, the MAE value of the algorithm UPCF is lower than the algorithm GSCF, which is lower than the UBCF and IBCF. This chapter shows that the error between the real score and the predicted score of the improved algorithm is the lowest, and the prediction of the user is more accurate. On the data sets train1 and test1, the RMSE values of the four recommendation algorithms are compared as shown in Table 2.

As can be seen from Table 2, the RMSE value of the algorithm UPCF is always smaller than the other three recommendation algorithms (under the same number of neighbors). As the number of neighbors increases, the RMSE value becomes smaller first and then bigger. When the number of neighbors is about 40, the value of RMSE tends to be the smallest, that is, the error of the prediction score is the smallest.

According to the data in Table 2, the RMSE values of the four contrast algorithms are plotted as histograms as shown in Fig. 5.

As can be seen from Fig. 5, with the increase of the number of nearest neighbors, the RMSE (root mean square error) value of the four algorithms is gradually reduced and then tends to be gentle. In the case of the same number of neighbors, the RMSE value of the UPCF algorithm has been lower than the other three kinds of recommendation algorithms, that is, the error between the real score and the prediction score of UPCF is the smallest, and the prediction is more accurate.

In order to exclude the impact of the data set on the results of the algorithm, the following experiments will be performed on second data sets train2 and test2 generated randomly.

In the data sets of train2 and test2, with the nearest neighbor as a variable, MAE and RMSE are the evaluation criteria to analyze and compare these four recommendation algorithms.

Figure 6 is the contrast effect diagram of the MAE value, with the increase in the number of nearest neighbors, the MAE value of the three algorithms decreases first and then increases. When the nearest neighbor number is about 40, the MAE value is the smallest, which shows that the collaborative filtering algorithm is affected by the nearest neighbor number. When the nearest neighbor number is the same, the MAE of UPCF algorithm is the smallest, that is, the predicted score is close to the real value of the user and it provides the best recommendation result. It fully illustrates the importance of the user to the subjective behavior of commodity score.

Figure 7 is the contrast effect diagram of the RMSE value, with the increase in the number of nearest neighbors, the value of RMSE decreased first and then increased. When the nearest neighbor number is about 40, the value of RMSE is the smallest. In the case of the same number of near neighbors, the RMSE value of the UPCF algorithm is the smallest, that is, the error between the real score and the prediction score is the smallest, and the performance is the best.

The recommendation algorithm GSCF based on graph structure and item type is based on the conventional algorithm, which makes full use of the indirect neighbor and commodity type when computing similarity.

However, the algorithm UPCF thinks that the user is the product of the user’s subjective behavior; it is a kind of implicit user preferences. We make full use of the user score behavior that can find the root cause of data sparsity, as it causes no marks of the product for users. In order to better analyze and compare the two improved algorithms in this paper, we will compare the MAE value and RMSE value of the two algorithms.

Figure 8 is the algorithm UPCF and algorithm GSCF, which has the contrast line chart of the data set train1 on the MAE value (the average absolute error). It can be seen from the graph that the MAE value of the two algorithms decreases first and then increases, and finally tends to be gentle. With the near neighbor number as the variable, When the near neighbor number is about 40, the MAE values of the two algorithms are the smallest, that is, the error is the smallest. Under the condition of the same neighborhood, the numerical MAE of the UPCF algorithm is lower than the GSCF algorithm. The MAE value is smaller, which means that the scoring error of UPCF is lower than GSCF, that is, the recommendation effect of the algorithm UPCF is more accurate than the GSCF algorithm. Because it analyzes the root causes of data sparsity, it makes full use of the implicit information implied by user rating behavior, namely, the user’s implicit preference.

Figure 9 is the RMSE (RMS error) contrast histogram between the algorithm GSCF and algorithm UPCF. The left is the algorithm GSCF, and the right is the algorithm UPCF. With the near neighbor number as the variable, the RMSE values of the two kinds of recommendation algorithms decrease first and then increase, and finally tend to be gentle. When the near neighbor number reaches about 40, the RMSE value of the two algorithms is the smallest, that is, the algorithm has the least error and the highest precision. In the case of the same number of near neighbors, the RMSE value of the algorithm UPCF is lower than that of the algorithm GSCF, that is, the error is smaller.

The experimental results show that the numerical values about MAE and RMSE of the UPCF algorithm are lower than the GSCF algorithm, namely, the error score is lower between the real and prediction score, so that the recommendation algorithm of UPCF is better than the GSCF algorithm.